Exclusive | SiliconFlow, the First to Catch DeepSeek's Traffic Wave, Raises Hundreds of Millions in New Funding Led by Alibaba

"A series of choices that were accidental, yet inevitable." By Lili Yu

Anflow Waves has learned that AI infrastructure company SiliconFlow recently completed a financing round of several hundred million RMB, led by Alibaba Cloud. Existing shareholders including Sinovation Ventures made additional investments beyond their pro-rata rights, with China Renaissance serving as the exclusive financial advisor. Earlier investors include Meituan (strategic investment), Sinovation Ventures, Glory Ventures, MiraclePlus, China Growth Capital, and Puhua Capital.

SiliconFlow founder Jinhui Yuan described this round as "a mutual sprint toward each other." Breaking this down: Alibaba has long treated AI infrastructure as a strategic priority. Earlier this year, Alibaba CEO Eddie Wu announced a massive investment of 380 billion RMB in cloud and AI hardware infrastructure — the largest investment of its kind by a Chinese private enterprise in history. From SiliconFlow's perspective, beyond the capital, "we'll be able to collaborate more deeply with Alibaba's Tongyi Qianwen ecosystem, and work together extensively on compute resources and domestic and international market expansion," Yuan said.

This also marks a post-explosive-growth financing for SiliconFlow. Its previous round came before the breakout, in late 2024, when it completed a Pre-A round of over 100 million RMB led by China Growth Capital, with Puhua Capital participating and existing investor Glory Ventures making an additional investment. This round's proceeds will be used primarily for talent recruitment, product R&D, and domestic and international market expansion.

Part01

Two Bets:

DeepSeek and Domestic Chips

SiliconFlow was a crucial piece of the miraculous year that DeepSeek brought about.

As the fastest To D (developer) and To B AI cloud service product to capture DeepSeek's traffic surge, its visit volume at one point spiked to surpass numerous To C applications.

Yuan noted that SiliconFlow's smooth adaptation of DeepSeek-R1 & V3 stemmed from the fact that "for a long time after DeepSeek-V2 was open-sourced, we were the only third-party MaaS platform supporting DeepSeek."

Tracing back further, this was because they bet on open source earlier. DeepSeek happened to be the open-source model whose exceptionally strong coding capabilities attracted widespread enthusiasm at the time.

Additionally, SiliconFlow's other major bet was on domestic chips. In early February, SiliconFlow partnered with Huawei Ascend to achieve efficient deployment of DeepSeek models on domestic chips.

As Yuan explained, the reason they were able to launch DeepSeek services based on domestic chips as a third-party platform ahead of others was that SiliconFlow had already done extensive R&D on domestic chips last year, solving numerous challenges in operators, communication, and system optimization. This allowed them to move faster than suppliers deploying DeepSeek on NVIDIA GPUs. Of course, the cost included that "our colleagues basically didn't rest throughout the Spring Festival holiday, though we paid triple overtime."

To this day, SiliconFlow remains the only supplier in the market offering large-scale DeepSeek API services on domestic chips.

Regarding recent progress on domestic chips, Yuan said that on one hand, domestic chips have become much easier to use, with no reliability issues. "In inference scenarios, because model architectures have converged, once you adapt the inference engine and deploy on a MaaS platform, developers can no longer perceive what underlying chips are being used." On the other hand, in practice, "domestic chips offer very high cost-performance for large-scale inference services."

Part02

After the Explosion:

What's Replicable and What's Not

The successful bets on open-source large models including DeepSeek and on domestic chips gave SiliconFlow growth far exceeding expectations. To date, SiliconFlow's total user base has exceeded 6 million, with thousands of enterprise users and daily token generation in the hundreds of billions.

Of course, DeepSeek's open-source strategy, while expanding the industry's scale, has also intensified competition among downstream MaaS providers.

Whether MaaS can be profitable has long been debated. Yuan believes MaaS is a product form that's user-friendly but challenging for suppliers.

In his view, to complete the commercial loop requires solving a series of technical, operational, supply chain optimization, and customer acquisition challenges. Solve these well, and MaaS can achieve a win-win for both users and suppliers. Currently, beyond exploring development paths domestically, they are actively expanding into overseas markets with better payment capabilities and industrial ecosystems.

Furthermore, he sees SiliconFlow's unique advantages as including: compute neutrality, model neutrality, and scenario neutrality.

Yuan is a serial entrepreneur. Before founding SiliconFlow, he invented LightLDA, then the world's fastest topic model training algorithm and system, at Microsoft Research Asia.

Prior to SiliconFlow, his entrepreneurial track record includes two other ventures: OneFlow, an open-source deep learning framework developed at OneFlow Technology, and the once-buzzy Light Year AI in earlier years.

In an interview with Anflow Waves last year, he said his first venture felt more like performance art: obsessed with technical pursuit, ultimately trapped by commercialization. Light Year AI, meanwhile, felt more like an accident that happened to everyone.

With this current venture, he feels he's finally developing a commercial touch, and the team "has gradually transformed from a lab-style organization into a more complete, mature commercial organization."

He attributes SiliconFlow's success to a series of correct choices and the team's execution: betting on open-source large models, focusing on AI inference services, and going all-in on adapting to domestic chips. "While explosive growth requires timing, making a series of correct choices is replicable and sustainable," he said.

Image source | IC Photo

Recommended Reading

Where Money Flows, People Rise and Fall